AI Revolution LIVE: Agentic AI App Building, Tech Earnings, And GPT-5

Something to think about here.

Hello and welcome to the AI revolution. Call. I'm Cody Willard, that's Bryce Smith. We work together. That's John dio. He's an AI professional for 30 years. We're here to talk about everything AI related and. Or AI investing. Investing in AI related.

Yes.

So Bryce, we have a few topics. I'm gonna just hit the. These are some of the questions and or framework for today's call. Is your. Oh, that's the title of the. The call. I guess we're gonna deliver Alpha, baby. We're not seeking alpha, we're delivering alpha. There's an AI web app for you, Bryce. Delivering Alpha. AI.

I'm gonna, I'll go fire up replit and, and have it built before the call's over.

What you can do, you could take it instead of having actual transcripts. It instantly finds the transcript when it gets published for each earnings report or investors relations event and summarizes it in AI. Creates graphs and mar. And AI web app related information off whatever data was given out on the report. I mean this, you can do that in real time. Bryce?

Yeah, I can show you. You want me to show you that Excel app that everyone's been talking about?

Sure.

So it's because it's sort of on this page but like can you clear.

Off this comment here? It's taking up my.

Oh my gosh, John.

Covering me up so there last year.

It's broken up. Funny. That's not going to be a normal. Okay, go ahead, go ahead, Bryce.

A couple weeks ago there was a. See here this.

Sorry, it's his first time on the Internet.

So these guys shortcut AI. They made their own version of Excel and it's AI native and you just talk to it and so like I, I just, I crudely pasted a half baked model I had for Google. This was my original data. Like you can see there's like formulas don't even work. Like I was like I'm just going to test this thing out, you know what I mean? Like here's some half baked model and I don't know if we can go find my original prompt on here, but I was like I want a, a comprehensive Google model with all of the pieces of the business broke out, right? Search YouTube, Google network hardware, the cloud, Waymo, right. And then build it out to 2040. Then I gave it some more details on that. But it's been like 15 minutes building this model and I've had to fix a few things. But like it put all of this Together and it has the formulas in here so I can go tweak, you know, if I don't like the growth rate, I can go change it to 80 or whatever. Right. But it's like it is all here. And I would think if, even if this isn't going to be perfect, if I was going to build a model on a company from scratch, this would still save me a lot of time.

Yeah, right. I mean you've done this. Right. Like when we put together a model for Tesla last year that went out 15 years.

Years.

I bet you spent two weeks.

Yeah.

Off and on every day for a few hours each day building the model, refining it to figuring out what other sectors you're missing, how you can improve it. And this would have cut that 90%.

Yeah. And I think what you realize is that like you spend so much time on the legwork of like, okay, this needs to multiply and then these need to add up and it needs to flow through to here. Right. Like that's sort of basically logic that you don't need a very analytical mind to do. Right. Like where our value add comes in is like, okay, what do we really think the YouTube growth rate is going to be? What's going to be the average time people spend on YouTube? Like that's where your thinking comes back in. But we don't need to be spending our time doing J8 times 365 divided by a thousand. Right. Like that sort of monotonous type of stuff. I do think the AIs obviously can do that fairly well. And then you can just come in here and change the inputs and make the model make sense. Like the model doesn't make sense right now. It's got, you know, $3 trillion of, $5 trillion of revenue by the year 2040. I think that's a bit optimistic. But.

It'S not impossible.

Not impossible, but like it's got some errors. Like so you know this Google. Oh, that's what I was trying to work on, the Google Network revenue. Like it didn't finish building that piece.

Out, but you can just tell it. Yeah, finish building out the.

I didn't understand what Google Network is. So you can see I said you need to build out the Google Network model. And I don't understand what that is. So Google Network is Google's like third party ads business, which includes things like AdSense and what do you call it, demand side platform, DSP and so on.

Enter understanding Read.

One of the things that is kind of cool when it does this is it usually comes back and asks you questions. I'll see if it does it.

Yeah.

Okay. Yeah. So a lot of the earlier it would ask me more questions like do you want these on separate tabs? Do you want it all on one tab? Like how long should the model be as it goes? 2040. Like, you know what I mean, it kind of iterates with you before it just gets to work. And anyway, this is a 200amonth product. I'm on a free trial. But you know, this is the, this is the most expensive tier you can get like a 20amonth version. But it's what I can do.

And would we call this vibe spreadsheeting?

Vibe modeling.

I like vibe spreadsheeting because anytime you get to say the word sheeting, it is funny.

Yeah, you're the only one. But you know, yeah, vibe modeling is good. So it's sounds expensive, but for what it does and for people who are heavy spreadsheet users sitting there all day in front of spreadsheets, it's, it's killer.

Yeah. But John, you tell me, I mean, how long before Google's sheets and Microsoft Excel are every bit as good as what these guys have built? Six months?

I don't know the Microsoft.

I mean, it's a fair question because the thing that their thing that they're doing is like horizontal. It's not like there was some choice to go into. Like Google would have to say, oh, we're going to get into, you know, retail laundromats or what. It's not like it's down some very specific thing. It's a thing they already do. Like sheets.

Well, they do spreadsheets and they do AI LLM, so. Pretty easy to marry the two.

Yeah, like there's, there's work there. These guys spent time making this. But I, you know, think Google has a few dollars in the bank to go do that. It's a, it's, it's more like just. Do they do it?

Yeah, they have to. I mean they, I, I don't. There's, that's prisoner's dilemma if nothing else. Bryce, what did you say about Microsoft? You don't think Microsoft will get this into Excel or you're being facetious.

A little tongue in cheek on that. But I do think like Microsoft has not really changed excel in 25 years and they've had plenty of opportunities to do so. And we're still manually typing in formulas and if then statements and pivot tables. Like it's just that institutional inertia is against Microsoft on changing Excel and It's a product that everyone knows how to use and so they have to go, you know, risk pissing off their users if they don't get the integration right.

You know, isn't that what's happened with Adobe Photoshop itself where they've, they have no choice but to keep adding AI technology into their flagship product? But like, existing users are getting confused and they don't remember how to use it. They don't know how to use Photoshop now, at least like they did.

So I wonder whether there's a decision that has to be made with these companies for how they structure things. Like, could Google go, yeah, we're not going to mess with sheets, we're just going to come out with the competitor shortcut right alongside of it or something. And you know, yeah, they both would go and make spreadsheets, but kind of making where you don't have to go affect existing users, make it as, as seamless as possible, not affect existing users.

Yeah, you know, you used to get the option don't upgrade, don't, you know, don't download the latest versions of like. But you know, somewhere along the app revolution in the last 10 or 15 years, you're like, I like your app. Like, I, I still have The Tiger Woods 2017 EA Go.

Don't even. Yeah, started.

Yeah, well, it doesn't work anymore. I paid for that thing.

So that the, in this case, because the, the, the kind of file you're making is the same, it's argument for, it's just in the same product. It's just a different one of the, it's like pivot tables, you know, it's just another thing it does. It's like, oh, you want to go on, on. It's, it's, it's, it's spreadsheet autopilot.

Microsoft does have Copilot built into Excel, but nobody uses it. Right. That's, that's why this thing went viral. And all the fintech financial meme handles on Twitter were going crazy and saying, you know, this is, you know, the, every analyst at Goldman Sachs is screwed type of thing.

I wonder also about like the stuff you can make with it. You're, you can do better with this. The more that you're a power user and understand what all of the potential models are for you to go put together. You know, if you're just the average user, it'll help. You know, you can get some things done you wouldn't have gotten done, but if you're a power user and you know what to ask for that's, I think, where it really pays off.

Yeah. There still is an art to the prompt engineering, or whatever you want to call it. You called it context engineering last week, John. But I know, like, even in replit, Cody and I have spent a lot of time in replit, and there are some tricks to getting replit to do what you want it to do.

That's another angle on it.

I'm yelling at replit right now on the other screen.

The. He did not finish.

All right, well, let's keep going. It's good stuff.

Yeah.

Johnny, couple other topics on the call. Today we will analyze and debate the agentic AI gold rush, moving beyond chatbots. Which companies are building the autonomous agents, autonomous AI agents that will dominate the new app economy. You know, John, that's what we're getting at. There is just this idea that, like most people's version of AI is still, hey, you talk to chat GPT or you Google something in CAI mode. And that's cool. Like, it's with, you know, the chat GPT moment two and a half, three years ago is all about that. But like, this AI app building phenomenon and being able to, like, any idea you can come up with for your phone or your computer, you can tell Replit or AI Studio or Bolt New or whatever, all of these new app builders, you can just build whatever's in your mind. And it's interesting that I think, Bryce, what you're getting at there is there's probably like two different. Three different ways even this all works for it. You have Nvidia down underneath it all. You got cloud providers from Google, Hyperscalers through Core Weave running the infrastructure. Underneath, you got the AI app builders, like the names. I just said Replit, John, Is it Replit. Replete. Bryce and I have, I think, Replit. Okay. Replit, Bolt, AI Studio, Price, say two or three others.

Yeah. There's Roark. Lovable. There's. There's also a layer in between that.

Wind surf.

Yeah, right.

Cursor.

Yeah, there's a layer in between the app builders and the infrastructure providers, which is the actual LLM builders too, because the reference aren't making their own LLMs.

Correct. So you. The cloud infrastructure to the LLM infrastructure to the. And by infrastructure on LLM, I mean the actual LLMs, this large language models like Anthropic Perplexity.

Well, there's becoming an infrastructure around those as you get to the, you know, agents and things, which is the other as. So we've been talking. We talked last week about, you know, this way to like rapidly, automatically, smartly create apps. But there were largely apps that were like the, you know, previous generation and then the next generation ones are the ones that are. They may be apps, but they, they are also smart. They're not smart to make them, they're smart in them. And that's where, you know, stuff we didn't talk about as much, but where things are headed with this, you know, these agent capabilities. It's like, well, what's that going to enable? The fact that you are making a thing that, you know, staying away from the AGI thing. But they start to act like what you'd expect. Like a person acts like you can ask it to do stuff. There's this now kind of way of measuring, of saying, like, how long can it go on its own? Used to be a minute, then it was an hour, now it's closer to a day.

What do you mean it can go.

How long.

It can process ideas on its own.

How long it can take one problem and go work on it. Yeah, like go, go off and like the research type stuff which both.

Open.

AI and Anthropic have, you know, where I think Google may as well where, you know, you kind of put it in research mode. And I did one that I fired off on ChatGPT and Anthropic at the same time with the same question. And the chatgpt one found like 15 references and the anthropic one found like 350 references. But the result on the chatgpt one was more on target to what I was looking for. So it was interesting that neither one of them has it totaled down yet, but both are interesting in. In what they can go do.

John, on that I have a question, like, there seems to be consensus now that Anthropic Claude is the best at coding. It's pretty much integrated into all of the AI coding applications as the default. You can most. Well, usually you can switch them out, but it seems like most people agree to Claude's best at coding. A lot of people think Gemini and Chat GPT or better like long form reasoning and analytics. People think Chad GBT is the best at like emotional, like a lot of people. Like, you know, all the influencers are like talking about using it for therapy and like how to deal with their girlfriend or whatever. You know, like, do you. It's surprising to me because they're just these probabilistic models trained on much the same data and yet they're even within the LLMs. There's like the, there's a Divergence into what people think they're best at. Brock is best at science and physics. You know, like, is that true or is that just our perception of.

I mean, I, I think it's true. First off, now, why or how they differ? You know, there, there is, there are differences in the training, there's differences in the, in the reinforcement learning. There's like differences like what they were even aiming at with some of the different, with some of the different ones. I've come to use Anthropic Claude as my go to today. I was doing something And I thought, ChatGPT can handle this. And it was a cipher. Query cipher is like SQL for graph databases.

SQL is you're going to one acronym to the other.

And SQL as most people who, like, who've been dealing with spreadsheets but then go on to like pivot tables, they get on to things like select employee from employees table where name starts with S or something. So that's SQL Structured Query Language. And it just pulls things out of the type of database people have been working with for the last 20, 30 years. So the newer kind of database, the graph database, has its own language like that, like Structured Query Language. So it's called cipher.

How do you spell cypher in this.

Case, C, Y, P, H, E, R. It's now an anti ansi standard, which means it's accepted as the way to do it. Just like SQL is for relational. I go to the language models and I say, make me this query because the query is kind of complex. You know, complex in code might be a couple hundred thousand lines, whatever. Complex in one of these queries means that it's, you know, 40 or 50 lines. It's this thing about yay big, but it's still got enough going on in it that, you know, I want help. So I get it and I'm running it and from Chat GPT and I'm getting the result back and I'm like, no, that's not right. And I'm showing it what's not right about it. It's like, oh, oh, that's because I did this. Oh, that's because I did that.

And.

And it starts correcting things and then when it gets around to it, it's fixed these things here and like whack a mole. The first thing it was got wrong is wrong again. So it's like, all right, let me just. So I go over to Anthropic and I hand it the query over. Late it was. And I go, here's something that doesn't Work and what I'm trying to get it to do is blah, blah, blah, make this work first try, just hand it back a working query. So there's definite difference.

It's incredible. Just even with you describing what you do every day that I use these AI app builders right now I'm just flabbergasted at the amount of code it can generate off of a random prompt. I was on Huggingface Co today just looking at the databases of AI applications and different, I don't know, whatever else they had on there. But, you know, one of the things, the most popular thing on there right now is a Deep Seek.

Deep Seek? Yeah.

Well, no, it's a Deep seek. Under Deep Seek is underlying. It's like deep code or something. And it's, it's an AI app builder built on the Deep Seek LLMs. And it's like, I didn't have to log in and I was like, hey, here's my idea. And it cranked out not a finished product, but 80, 90% of what was a finished product in about five minutes. And this is again a free website that I didn't log into using a Chinese LLM that I didn't get log into, none of which is running on anywhere near my computers. And I'm like making web apps that actually do stuff.

The seems like the, the limitation is just what can you think of doing? Because on the side of doing it, that's not the limitation.

So, John, you know, like Bryce, you're making this like eating calorie tracking fitness app that you had in your idea brain. If I may. Basically, it's like you were using one. You're like, this thing sort of sucks. I should be able to like, take a picture. It should know what's in, like the caloric content, at least generally, of whatever I'm about to eat. But even the thing that you're doing right now, it works. You take a picture of a hamburger and it's like, hey, that's a hamburger. It's generally 550 calories, unless it's got cheese and then it's 650. And the fat contents this. It gives you all your data. It's fantastic. You can submit. I'm tracking this. No, no. But it will get to the point where you're literally just wearing your meta glasses or some button on your shirt or something. And it knows that you're sitting down to eat, it knows that you're drinking something, it knows that you're even eating a granola bar on walking. And it's like, oh, snapshot. Oh, that was a Quaker Oats chocolate chip bar. That's 280 calories. And you're not even going to have to engage with it. It will be passively identifying itself around your daily health, eating and workout habits. We talked before already sort of knows that, right? Like when I, when I go on a bike ride, it's, it's a Tim Cook Apple technology. So it's not very good, but it at least about half the time it's like, hey, you might be 10 minutes into a bike ride. Should I track that? I'm like, yeah. And it's like, all right, you're three minutes into the bike ride.

Okay, thanks. But.

We'Ve talked about like, well, what's missing from these LLMs or you know, capability wise to get where we're going and like one of the things we've talked about is memory and how. So when you sit there at a chat prompt and you ask something and you get a really smart answer and a second later you ask something and it shows it has no idea that you asked something a second ago or what that had to do with something. I'm picturing Bryce with his app and is, you know, taking a picture of food or, or just it's on, you know, video. Like you're saying it's the glasses can see it or whatever for it to like keep track over time and go, you know, that's, that's every day this week that you've blah, blah, blah or you know, and be able to start to, to deal with the pieces and piece them together. And it seems to me that most of what's needed has been invented. It's a matter of like putting it together.

Like it's. We need compute and memory. Like we need a thousand times more compute and a thousand times more memory capabilities and then, you know, then the cost of remembering your food for a week is truly de minimis in the cloud.

And well already if it's able to turn the video at you know, 1 gigabyte per whatever into a little phrase of text, you know, granola bar. And it can put, it can say granola bar a million times right now on your, you know, it's nothing.

Yeah, yeah. I think that's what makes rep. What replit did so well is a. Basically they integrated a bunch of stuff that already existed in a really intuitive way. So that way as a non developer you can make apps. You know, databases existed and authentication existed and LLMs existed and it's like, hey, let's figure out A way to make these things talk to each other so an LLM can build it. And that's basically what replit is. They didn't have to go out and build their own LLM or database or anything. They just integrated what was all there and it just works.

So again it's down to the idea. It's what are, you know, what are you going to come up with to make and then how are you going to distribute it? How are you going to get feedback whether it is doing what people need or not. And then how are you going to set that up to have a good loop back on? Oh good, I'll make that update. You know, if you're wanting to do something like out there with what you're doing, it's. It's up to the ideas now and carrying them out.

Absolutely, that's right.

So let's get next topic. That's what we just said. Agentic gold. Actually John, on that note, are we going to get to the point that we're envisioning where you can even just talk through the app to some interface and you're not even looking at a computer screen and correcting it and it'll just. Yeah, I need to live stream that and I'll put that feature in for you.

And so thinking two pieces that have to happen for that. One is just the voice interface which clearly exists in other places, but now even I think in the latest version of Windsurf, which is one of these automatic coding apps, it's one of the type that it's full, you know, full stack, it'll make front end and back end code for you. So they just incorporated voice so you can sit there and talk in what it is you want to have happen. But the other parts, the other thing you said there is you've gotta like lower the error rate enough that you're not bringing them back, you know, to the screen for that. Or is there a way, you know, it's not done until it's done just to do with audio in the other direction of okay, I got what you're saying, but can you tell me this or whatever. So yeah, it's one of those ones where I don't think it needs new invention of anything. Somebody just needs to go do it.

Bryce, thoughts?

Yeah, I'm thinking a lot about the investment implications of this and Cody and I talked about this extensively, but it seems like to me that the companies that are providing some of the back end services, the pipelines, the databases, the things that you're still going to need in the AI app revolution, you know, even if the amount of apps that goes up by 100x or something, I think, I don't know that all of those companies are set up to benefit. You know, it just depends on how fast they can integrate. But like the Mongo databases or Okta, like these companies that already have existing services which you need in software development, you know, the faster that they can integrate into these AI tools and these AI app builders like there, there's going to be like, I think there's going to be a divergence where you have the ones that, that successfully pulled that off in the amount of time that they had and the ones that are, you know, still trying to find their way in this AI world, like they're probably going to get left behind, you know.

Well, there's, there's this kind of, that's, I don't have to call it middleware, but other pieces of the puzzle that are starting to fit in to make development happen faster. Like MCP model context protocol from Anthropic. But it is open source, free for everybody to go incorporate. And it's this way to make it easier to front end. So you got something using, doing something with an LLM and figuring out answers to things and even carrying out application things. It's the new kind of API in a way, except you don't have to go write code every time you plug it in. Then it makes your stuff smarter. There's things like that showing up to make this, to make there be new pieces to the puzzle, to make it easier to make the new stuff.

Is MCP the standard? Is that now the de facto standard? At one point there was sort of like, I thought Google had their own thing. Maybe Salesforce is using something else.

Google has its own thing which is a to a or agent to agent interface or something. And so MCP and the Google thing are complementary. You could, you could use them both. And to say standard would be a little, it's a little early for that even to say de facto standard, but widely used, you know, and people, you know, talking about it a lot and saying how they're making stuff happen with it. They're taking kind of a role like what Python did with programming where it's like just okay, everybody's using it for building their machine learning models because of the Pytorch library, just a thing that Python has available with it. Everyone's doing that now. Since everybody's on that bandwagon, new stuff just shows up there. And similar thing with, with MCP and, and with the A to A or whatever it's the other one's called.

Interesting.

John, is it possible, like, are these, these AI app builders that we're talking about, Replit, Bolt, AI Studio, etc, do they. And now Microsoft this week, by the way, released their Copilot AI app builder. Also, I'm just trying to sort of like, figure out, looking three or five years from now, are one or two of these truly differentiated? Are they commoditized? Because it feels like there's some commoditization of it already when we're like.